The Usage of LLM in Grab & Gojek
How Southeast Asia's Leading Super Apps are leveraging Large Language Models to enhance their platforms.
Introduction
Grab & Gojek are two of the biggest super apps in Southeast Asia. The app includes ride-hailing, food delivery, digital payment and many more. Both companies are leveraging Large Language Models (LLMs) to enhance their platforms but in distinctly different ways.
Grab and OpenAI
Grab has collaborated with OpenAI to integrate GPT-4's vision fine-tuning to allow GrabMaps to localize speed limit signs, turn restrictions, places, and road geometries more accurately.
- Speed limit sign localization: The initial use case involved matching speed limit signs to their respective roads.
- Turn restriction and geometry identification: The model identifies intricate details like elevated roads and turn restrictions, which are crucial for navigation.
By using their network of motorbike drivers and pedestrian partners, equipped with 360-degree cameras they have collected millions of street-level images. With the street data collected and by using OpenAI's vision fine-tuning, Grab has made mapmaking faster, smarter, and more efficient. These highly detailed maps improve navigation for millions of users and driver-partners daily, boosting economic activity across Southeast Asia.
To find out more about the collaboration visit: OpenAI x Grab
GoTo, Indosat Ooredoo and AI Singapore
PT GoTo Gojek Tokopedia Tbk has collaborated with Indosat and Indosat Ooredoo Hutchison, in partnership with AI Singapore, have launched an open-source Indonesian-focused language model called Sahabat AI.
Model Size
They chose smaller models (7B, 8B, and 9B parameters) instead of larger ones, prioritizing efficiency. These models were fine-tuned from AI Singapore's SEA-LION v2, which supports Southeast Asian languages like Thai, Vietnamese, and Tamil.
Cultural Focus
Sahabat-AI specializes in Indonesian language and dialects. It was trained on:
- 448,000 Indonesian instruction-completion pairs
- 96,000 pairs in Javanese
- 98,000 pairs in Sundanese
- 129,000 pairs in English for additional versatility
By being open-source, the project seeks to make large language models (LLMs) more accessible to businesses and individuals in Indonesia and further enhance the model capabilities and relevance.
To find out more about the model visit: GoTo Company on Hugging Face